Gemini 3.5 Flash integrates native computer use tool

Google integrated computer use directly into Gemini 3.5 Flash, allowing agents to observe screens and execute actions across browser, mobile, and desktop surfaces. The change eliminates the need for a separate model call. Developers access the capability through the existing Gemini API and the Gemini Enterprise Agent Platform. Targeted adversarial training was applied during fine-tuning to reduce prompt injection success rates in live environments.

The release adds two optional enterprise controls: mandatory user confirmation before sensitive actions and automatic task termination on detected indirect prompt injection. These sit alongside existing recommendations for sandboxing and access controls. No quantitative benchmark scores for computer use accuracy or step completion were released in the announcement. Prior standalone 2.5 model results are not restated for the integrated version.

Operationally the update shortens the path from prototype to production for repetitive knowledge-work and test-automation agents. It aligns Gemini with Anthropic’s computer-use API and OpenAI’s operator prototypes, though each vendor retains distinct safety layers. Production workloads remain gated by human-in-the-loop checkpoints and sandbox requirements.

Next measurable signal will be the count of production agents calling the computer-use tool through the Gemini API after the first full quarter of availability.